Markets
OpenLake Posted 1.98x NVIDIA AIStore's Write Bandwidth on MLPerf Storage v3.0 Checkpointing
OpenLake published MLPerf Storage v3.0 results on 2026-09-01 and took them to Show HN on 2026-09-05 (35 points). On the Llama 3.1 8B checkpointing workload it recorded 6.72 GiB/s write and 11.55 GiB/s read against five identically configured submissions, beating NVIDIA AIStore's best 6-node result of 3.40 GiB/s write and Nebius Object Storage at 2.81 GiB/s. Read-side margins are thin (11.55 vs 11.08 GiB/s), so the claim is specifically about checkpoint write throughput, which is the part that stalls training loops.
↳ Follow the thread